451787 - TM: Trace unary not for numbers and strings

Status: RESOLVED FIXED

(Core :: JavaScript Engine, enhancement, P3)

Description

[Boris Zbarsky [:bzbarsky]]

TraceRecorder::record_JSOP_NOT bails out of the arg is not boolean or object.  I don't see why we couldn't use the code it has right now for INT and DOUBLE as well...

As far as that goes, we should be able to do strings too, right?

Comment 1

[Andreas Gal :gal]

Yeah, we just add stuff on demand as we go along (gives an entirely new meaning to incremental compilation).

Comment 2

[Mike Shaver (:shaver -- probably not reading bugmail closely)]

For double, eq0 isn't sufficient, since NaN != 0 but !NaN is true.  I bet we can get there by chaining enough eq0s together, though. :)

Comment 3

[Boris Zbarsky [:bzbarsky]]

So I tried just doing that, but I ran into the little issue where Assembler::asm_cmp has this to say:

537         // Not supported yet.
538 #if !defined NANOJIT_64BIT
539         NanoAssert(!lhs->isQuad() && !rhs->isQuad());
540 #endif

Comment 4

[Boris Zbarsky [:bzbarsky]]

Attachment: Simple JS test file

This tests the performance (delta1 and delta2 should be about equal, not differ by a factor of 8) and the NaN behavior (i should end up as 1000).

Comment 5

[Blake Kaplan (:mrbkap) (inactive)]

Boris, the parser will nuke useless if statements like the one in your first test, so the first test is equivalent to |for (lots of stuff) /* empty */;|

Comment 6

[Boris Zbarsky [:bzbarsky]]

Ah.  I guess I should assign the false to a var.  In any case, the second test should be about as fast as the first one.

Comment 7

[Andreas Gal :gal]

Attachment: Add support for !string.

Comment 8

[Andreas Gal :gal]

http://hg.mozilla.org/tracemonkey/rev/4d02662a0cd2

Comment 9

[Bob Clary [:bc] (inactive)]

test also included in js1_8_1/trace/trace-test.js